HW config

HW Parameter Value
0 No of cards 8
1 No of tiles per card 2
2 Batch size 32
3 Network Topology p2p
4 Frequency 1.5
5 Scale up option GLUELESS
6 Serdes rate (Gbps) 90
7 Buffers Enabled
8 Buffer size 104857600
9 No of PVC 128
10 No of tiles per PVC 2
11 No of PVC per host 8
12 Bw per NIC unidirection (Gbps) 400

SpeedSim Task analysis

Fabsim Task analysis

Run Summary

Metric Speedsim Fabsim
0 Total Compute time per tile(ms) 217.323 217.323
1 Scale Up no overlap time (ms) 0.222439 0.025508
2 Scale Up overlap with Compute, per tile(ms) 0.222439 0.025508
3 Scale Out FW time (ms) 0 0
4 Scale Out No overlap, Model parallel (INP) time, per tile(ms) 0 0
5 Scale Out overlap with Compute, Model parallel (INP) time, per tile(ms) 0 0
6 Scale Out No overlap, Data parallel (WT) time, per tile(ms) 0.306507 0.220237
7 Scale Out overlap with Compute, Data parallel (WT) time, per tile(ms) 0.306507 0.220237
8 Total Scaleup comm overlap with Scaleout comm and not with Compute (ms) 0 0
9 Total time without overlap per tile (ms) 217.852 217.569
10 Total time with overlap (Compute with Comms) per tile(ms) 217.852 217.569
11 Scaling efficiency (%) 99.7572 99.887
12 Throughput compute only per tile 147.246 147.246
13 Throughput full overlap per tile 146.889 147.08
14 Throughput no overlap per tile 146.889 147.08
15 Throughput full overlap 37603.5 37652.5